Image-based classification of wheat spikes by glume pubescence using convolutional neural networks

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Morphology

Background:

• Pubescence, a key plant trait, enhances resistance to environmental stresses like drought and pests.
• In wheat, glume pubescence is a critical morphological marker for variety classification.
• Current methods for determining pubescence are subjective or labor-intensive.

Purpose of the Study:

• To develop an automated, objective method for determining wheat glume pubescence.
• To leverage digital imaging and convolutional neural networks (CNNs) for this task.
• To improve the selection of stress-resistant wheat cultivars.

Main Methods:

• Image segmentation using U-Net with EfficientNet-B1 encoder to isolate wheat spikes.
• Classification of glume pubescence (pubescent/glabrous) using CNN architectures (Resnet-18, EfficientNet-B0, EfficientNet-B1).
• Training and testing on a dataset of 9,719 spike images.

Main Results:

• The U-Net model achieved high segmentation accuracy (IoU = 0.947 for spike body).
• EfficientNet-B1 demonstrated superior classification performance (Test: F1 = 0.85, AUC = 0.96; Holdout: F1 = 0.84, AUC = 0.89).
• Higher image magnification and reduced distortions improved prediction accuracy.

Conclusions:

• An integrated approach using CNNs effectively automates glume pubescence determination.
• This AI-driven method offers an objective and efficient alternative to manual classification.
• The findings support the use of AI in plant breeding for identifying desirable traits like stress resistance.